DianShi-RetroPlanSystem-Assets

Released with the paper Reaction precedents extend retrosynthetic prediction beyond training coverage.

Runtime assets and benchmark targets for the three multi-step planning frameworks in DianShi-RetroStepLM: RetroAgent, Retro-R1-modified and DianShi-RetroPlanSystem-v0. The single-step model they call is DianShiGroup/DianShi-RetroStepLM-Direct.

Contents

Path	Size	Description
stock/emolecules_standardized.sqlite	1.1 GB	23,081,629 RDKit-standardized purchasable building blocks from eMolecules. Table stocks(smiles TEXT PRIMARY KEY).
rdchiral_templates/DianShi-PatentRxnStep-Train-Full_rdchiral_templates.jsonl	88 MB	412,027 rdchiral reaction templates extracted from DianShi-PatentRxnStep-Train-Full, used by the template gate, one per line: {"key": "<template>", "value": <count>}.
test_dataset/*.txt	<1 MB	Benchmark targets, one SMILES per line.
test_dataset/*_DATASET.yaml	—	Per-set metadata: case count, visibility, sha256.
SHA256SUMS	—	Checksums for every file.

The stock comes from the Retro* eMolecules building-block list, standardized with RDKit 2025.9.5, and then converted to SQLite.

SCScore checkpoints are vendored in the framework repositories, not in this dataset: RetroAgent/datasets/scscore/ contains the 1024-bit tool-server checkpoint, and each framework's gate directory contains the 2048-bit checkpoint.

Download

Download into the framework directory (multi_step/<framework>/ in DianShi-RetroStepLM):

hf download DianShiGroup/DianShi-RetroPlanSystem-Assets --repo-type dataset \
  --local-dir DianShi-RetroPlanSystem-Assets

# Skip the 1.1 GB stock:
hf download DianShiGroup/DianShi-RetroPlanSystem-Assets --repo-type dataset \
  --local-dir DianShi-RetroPlanSystem-Assets --exclude "stock/*"

Verify with cd DianShi-RetroPlanSystem-Assets && sha256sum -c SHA256SUMS.

Usage

The frameworks' .env.example files and benchmark configs already point at DianShi-RetroPlanSystem-Assets/, so no renaming is needed:

DIANSHI_RDCHIRAL_TEMPLATE_JSONL=DianShi-RetroPlanSystem-Assets/rdchiral_templates/DianShi-PatentRxnStep-Train-Full_rdchiral_templates.jsonl
STARTING_MOLECULES=DianShi-RetroPlanSystem-Assets/stock/emolecules_standardized.sqlite

Pass a target file with --targets (RetroAgent, Retro-R1 runner) or --data_file (Retro-R1 cli.py); DianShi-RetroPlanSystem-v0's benchmark configs read DianShi-RetroPlanSystem-Assets/test_dataset/ directly.

RetroAgent also loads the stock from a fixed path, so copy it to RetroAgent/dataset/emolecules_standardized.sqlite. The 1024-bit SCScore checkpoint is vendored at RetroAgent/datasets/scscore/; no copy is needed.

Benchmarks

File	Targets	Source
test_dataset/uspto_190_targets_smiles.txt	190	USPTO-190, introduced with Retro* (Chen et al., 2020)
test_dataset/pistachio_hard_tree.txt	100	Pistachio Hard (Yu et al., 2024)
test_dataset/patmol_150_targets.txt	150	PatMol-150, built for this study from 2025 public patents; reference routes 7–13 steps (mean 8.17)

Sources and licenses

The dataset mixes files we built with third-party data, so the license field is other. Each file's terms:

Files	License
rdchiral_templates/ — templates extracted with RDChiral from DianShi-PatentRxnStep-Train-Full	Apache-2.0
test_dataset/patmol_150_* — PatMol-150, built for this study	Apache-2.0
stock/ — eMolecules raw materials (https://www.emolecules.com), via the Retro* list	eMolecules' terms; check them before redistributing
test_dataset/uspto_190_*, test_dataset/pistachio_hard_* — public benchmarks	Original terms; cite the original works

Citation

If you use these assets, please cite:

@misc{sun2026reactionprecedents,
  title         = {Reaction precedents extend retrosynthetic prediction beyond training coverage},
  author        = {Jiaxing Sun and Xuli Shen and Jiang Wu and Ruiyi Ding and Ruikang Zhang and Xiao An and Xibao Li and Yubin Wang and Xingjian Wei and Yinfan Wang and Yu Liu and Lingli Ge and Xi Xue and Xiaojian Wang and Lijun Wu and Chang-Yu Hsieh and Yuqiang Li and Bowen Zhou and Dahua Lin and Conghui He},
  year          = {2026},
  eprint        = {2610.40249},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  doi           = {10.48550/arXiv.2610.40249},
  url           = {https://arxiv.org/abs/2610.40249}
}

For USPTO-190 and Pistachio Hard, please also cite the original works.